Con-Text: Text Detection Using Background Connectivity for Fine-Grained Object Classification

Sezer Karaoglu, Jan van Gemert, Theo Gevers

1

Can we achieve a better object recognition with the help of scene-text?

2

Goal

• Exploit hidden details by text in the scene to improve visual classification of very similar instances.

Application : Linking images from Google street view to textual business inforation as e.g. the Yellow pages, Geo-referencing, Information retrieval

3

SKYSKYSKY

CAR CAR

DJ SUBS Breakfast Starbucks Coffee Starbucks Coffee

Challenges of Text Detection in Natural Scene Images

o Lightingo Surface Reflectionso Unknown backgroundo Non-Planar objectso Unknown Text Fonto Unknown Text Sizeo Blur

4

Literature Review Text Detection

• Texture Based: Wang et al. “End-to-End Scene Text Recognition”

ICCV ‘11

Computational ComplexityDataset specific

Do not rely on heuristic rules

• Region Based: Epshtein et al. “Detecting Text in Natural Scenes

with Stroke Width Transform ” CVPR ‘10

Hard to define connectivity Segmentation helps to improve ocr performance

5

Motivation to remove background for Text Detection

• To reduce majority of image regions for further processes.• To reduce false positives caused by text like image regions (fences, bricks,

windows, and vegetation).• To reduce dependency on text style.

7

Automatic BG seed selection BG reconstructionText detection by BG

substraction

Proposed Text Detection Method

Background Seed Selection

• Color, contrast and objectness responses are used as feature.• Random Forest classifier with 100 trees based on out-of-bag error are used to create forest.• Each tree is constructed with three random features.• The splitting of the nodes is made based on GINI criterion.

Original Image Color Boosting Contrast Objectness

Conditional Dilation for BG connectivity

where B is the structring element (3 by-3 square), M is the binary image where bg seeds are ones and X is the gray level input image

untilrepeat

Text Recognition Experiments

• ICDAR’03 Dataset with 251 test images, 5370 characters, 1106 words.

10

ICDAR 2003 Dataset Char. Recognition Results

11

Method Cl. Rate (%)ABBYY 36Karaoglu et. al. 62Proposed 63

The proposed system removes 87% of the non-text regions where on average 91% of the test set contains non-text regions. It retains approximately %98 of text regions.

ImageNet Dataset

• ImageNet building and place of business dataset ( 24255 images 28 classes, largest dataset ever used for scene tekst recognition)

• The images do not necessarily contain scene text.• Visual features : 4000 visual words, standard gray SIFT only.• Text features: Bag-of-bigrams , ocr results obtained for each image

in the dataset.• 3 repeats, to compute standard deviations in Avg. Precision.• Histogram Intersection Kernel in libsvm.• Text only, Visual only and Fused results are compared.

Steak PizzeriaFuneralBakery Discount HouseCountry House

Fine-Grained Building Classification Results

ocr : 15.6 ± 0.4 Bow : 32.9 ± 1.7

Text Visual Fusion

Bow + ocr : 39.0 ± 2.6

#269 #431 #584 #2752

#1 #4 #5 #8

Visual

Text

Proposed

Discount House

#1 #4 #5 #8

Conclusion

• Background removal is a suitable approach for scene text detection • A new text detection method, using background connectivity and, color, contrast and objectness cues is proposed.• Improved performance to scene text recognition. • Improved Fine-Grained Object Classification performance with visual and scene text information fusion.

14

DEMO

TRY HERE